Bandwidth extension is a critical task in speech processing that enhances narrowband audio quality by reconstructing essential high-frequency components, which are vital for natural and intelligible speech. Recent advancements like EBEN (Extreme Bandwidth Extension Network) achieve notable results with lightweight generators and real-time operations. However, down-sampling limitations degrade frequency-space information, causing speech distortion as up-sampling struggles to recover lost details. To address this, we propose replacing the generator with an ICCRN in the frequency domain. Leveraging cepstral-domain features and inplace convolutions, our approach mitigates information loss during down-sampling and preserves global spectral structure. Evaluations on the French LibriSpeech dataset show significant improvements, including clearer spectrograms, fewer generator parameters, and lower computational costs, advancing the state of the art in bandwidth extension.